Multi-Vector Retrieval Strategies
What you'll learn: How to use multiple embeddings per document (summary, questions, content) to dramatically improve retrieval accuracy in RAG systems.
The Core Idea
Instead of creating just one embedding vector per document chunk, you generate multiple vectors representing different aspects of the same content. For example, you might create:
- An embedding of the full document text
- An embedding of a generated summary
- Embeddings of potential questions the document answers
- An embedding of key concepts or entities
When a user query comes in, you search across all these vector types. This increases your chances of finding relevant documents because different vector representations capture different semantic angles.
Why Multiple Vectors Work
The analogy: Think of a library book with multiple catalog entries. One entry lists it under "Biology," another under "Evolution," and a third under "Darwin." A researcher searching any of those terms will find the book.
Similarly, a document about "machine learning model deployment" might not match well with the query "how to serve predictions in production" using just the raw content embedding. But if you generated a question vector like "How do you deploy ML models to production?" during indexing, that query-to-question match would be excellent.
Implementation Pattern
At index time, for each document:
- Embed the original content
- Generate a summary and embed it
- Generate 3-5 potential questions it answers and embed each
- Store all vectors, linked to the same source document
At query time, search across all vector types and merge results (often using techniques like Reciprocal Rank Fusion you've already learned).
Key Takeaway: Multi-vector retrieval trades higher storage and indexing costs for significantly better recall by representing each document through multiple semantic lenses.